HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation

Learning, selection and coding of new block transforms in and for the optimization loop of video coders

Abstract : Transforms are a key element in block-based video coding systems which, in conjugation with quantization, is important for the overall compression efficiency of the system. This thesis explores multiple transform- based learning schemes. A first contribution of this work is dedicated to the evaluation of transform learning schemes with two flavors 1) online learning, and 2) offline learning. The two approaches are compared against each other and their respective appropriability is studied in detail. Some novel techniques are proposed in this work to 1) improve the stability of the learning scheme and 2) to reduce the signaling cost. In a second contribution of this thesis, the offline multiple-transform learning schemes already known in the literature are further extended with the aims to altogether 1) provide more generic transforms that are less biased towards specific classes of contents, 2) achieve higher compression gains, 3) reduce encoding and decoding computational complexity. An improved Mode Dependent Transform Competition (IMDTC) scheme is proposed which provides a considerable gain of over 5% compared to standard HEVC under All Intra (AI) configuration at a complexity just 2.9 times the standard HEVC. Finally, the content adaptability aspect of the offline learning is explored through a novel content-adapted pool-based transform learning approach where several multiple-transform sets are learned on different contents and pooled together. During the coding of a given region of an image, one transform set is selected locally from the pool. Numerical results show the high potential of this approach against the conservative online and offline approaches.
Complete list of metadata

Cited literature [106 references]  Display  Hide  Download

Contributor : Saurabh Puri Connect in order to contact the contributor
Submitted on : Thursday, April 26, 2018 - 4:39:27 PM
Last modification on : Wednesday, April 27, 2022 - 4:23:54 AM
Long-term archiving on: : Tuesday, September 25, 2018 - 10:23:07 AM


Files produced by the author(s)


  • HAL Id : tel-01779566, version 1


Saurabh Puri. Learning, selection and coding of new block transforms in and for the optimization loop of video coders. Computer Science [cs]. Université Bretagne Loire; Université de Nantes; LS2N, Université de Nantes, 2017. English. ⟨tel-01779566⟩



Record views


Files downloads